AI models like the DALL-E 2 continue to make art that looks very European

In late September, OpenAI made its DALL-E 2 AI technical generator at scale publicly available, allowing anyone with a computer to make one of those amazing and slightly weird photos that seem to be floating around the internet more and more these days. DALL-E 2 is by no means the first AI art constructor to be opened to the public (artifact models competing with artificial intelligence). stable spread And the Midjourney It was also launched this year), but it comes with strong lineage: its cousin, the text generation model known as GPT-3 – itself the subject of much intrigue and multiple intruder Stories – Also developed by OpenAI.

Last week, Microsoft announce It will add AI-generated art tools — powered by DALL-E 2 — to its suite of Office programs, and in June DALL-E 2 was used to design Cosmopolitan magazine cover. The most utopian proponent of AI-generated art argues that it provides the democratization of art to the masses. The cynics among us might argue that he is copying human artists and threatening to end their careers. Either way, it seems clear that the art of AI exists, and its potential is just beginning to be explored.

Naturally, I decided to give it a try.

As I scrolled through examples of DALL-E’s work for inspiration (I had decided my first attempt had to be a masterpiece), it seemed to me that the AI-generated art didn’t have any particular aesthetic other than, perhaps, being a little weird. There were pigs in sunglasses and flowered T-shirts while riding motorbikes, raccoons playing tennis, and Johannes Vermeer The girl with the pearl earring, modified slightly so that it replaces the titular girl with a sea otter. But as I continued to scroll, I realized that there was one unifying theme behind each piece: AI art, for the most part, resembles Western art.

“All AI is only looking back,” he said. Amelia Winger-BerskinProfessor of Artificial Intelligence and the Arts at the University of Florida’s Digital Worlds Institute. “They can only look into the past, and then they can predict the future.”

For an AI model (also known as an algorithm), the past is the data set on which it was trained. For an AI technical model, this data set is an art. A large part of the world of fine art is dominated by white Western artists. This leads to AI-generated images that look overwhelmingly Western. This is, frankly, a little disappointing: AI-generated art, in theory, could be an incredibly useful tool for imagining a fairer view of art that looks very different from what we’ve come to take for granted. Instead, it simply perpetuates the colonial ideas that drive our understanding of art today.

To be clear, models like the DALL-E 2 can be asked to create art in any artist’s style; Requesting a photo with the “Ukiyo-e” mod, for example, will create works that mimic Japanese prints and woodblocks. But users should include those rates; They are rarely the default.

DALL-E 2 interpretation of the instant “Hokusai Artificial Intelligence Painting”
Neil Danesha/Fox; Courtesy of OpenAI

Winger-Bearskin has seen the limits of AI art firsthand. When one of her students used the images generated by Stable Diffusion to create a video of her nature scene, she realized that the twilight backgrounds set by the AI ​​model looked strangely similar to scenes drawn by Disney animators in the 1950s and 1960s – which they themselves were. Probably by the French Rococo movement. “There are a lot of Disney movies, and what he got was something we see a lot of,” Winger-Bearskin told Recode. “There are a lot of things missing in these data sets. There are millions of night scenes from all over the world that we will never see.”

The bias of artificial intelligence is a Known problem with difficulty. Left unchecked, algorithms Racial and gender bias can be perpetuatedAnd this bias extends to the art of artificial intelligence as well: like Sigal Samuel Books in Future Perfect In April, previous editions of DALL-E would post pictures of white men when asked to portray lawyers, for example, and portray all the flight attendants as women. It was OpenAI a job to mitigate these effects, and adjust its model to try to eliminate stereotypes, although researchers still disagree on whether these measures worked.

But even if they succeed, the problem of artistic style will persist: if DALL-E succeeds in portraying a world free of racial and sexist stereotypes, it will still do so in the image of the West.

“You can’t set the model to be less western if the data set is mostly western,” Yilon DoA doctoral student and researcher in artificial intelligence at the Massachusetts Institute of Technology told Recode. The AI ​​models are trained by stripping the internet of images, and Du believes the models made by US or European-based groups are likely geared towards Western media. Some models made outside the United States, such as ERNIE-ViLG, developed by Chinese technology company Baidu, do a better job of producing images that are more culturally relevant to their place of origin, but they come with issues of their own; Such as MIT Technology Review reported In September, ERNIE-ViLG was better at producing animation art than DALL-E 2 but refuses to take pictures of Tiananmen Square.

Since the AI ​​is looking back, it is only able to make different variations of the images it has seen before. This, Doe says, is why an AI model can’t generate an image of a plate placed on top of a fork, even though it must understand every aspect of the order. The model simply didn’t see a picture of a plate on top of the fork, so it spits out pictures of the forks on top of the plates instead.

Incorporating more non-Western art into an existing data set would not be a very useful solution either, due to the overwhelming prevalence of Western art on the Internet. “It’s a bit like giving clean water to a tree that has been fed polluted water for the past 25 years,” Wenger-Perskin said. “Even if the water gets better water now, the fruits of that tree are still contaminated. Running the same model with new training data doesn’t change it significantly.”

Instead, creating a better, more representative model of AI requires building it from scratch — which Winger-Bearskin, a member of the Seneca-Cayuga Nation of Oklahoma and an artist herself, does when she uses AI to create art around the climate crisis.

This is a time consuming process. “The hardest thing is to configure the dataset,” Doe said. It would take millions of images to train an AI art creator, and Doe said it would take months to create a data set that is equally representative of all the art styles that can be found around the world.

If there is a silver lining to the artistic bias inherent in most art models of AI, it might be this: Like all good art, it reveals something about our society. Winger-Bearskin said that many modern art museums offer more space for art made by people from underrepresented societies than they did in the past. But this art still constitutes a small part of what is in the archives of museums.

“The artist’s job is to talk about what’s happening in the world, to amplify issues so that we can notice,” he said. Jin ohD., associate research professor at the Institute for Robotics at Carnegie Mellon University. The art models of AI are unable to provide commentary of their own – everything it produces is at the behest of a human – but the art it produces creates the kind of occasional meta-comment that Oh thinks is noteworthy. “It gives us a way to observe the world the way it is organized, not the ideal world we want it to be.”

That’s not to say Oh thinks that fairer models shouldn’t be created – they are important for circumstances where depicting an ideal world is useful, as for children’s books or commercial applications, she told Recode – but rather that the existence of imperfect models should make us think More in depth how to use it. Rather than just trying to eliminate biases as if they don’t exist, Oh said, we should take the time to identify and measure them in order to have constructive discussions about their effects and how to reduce them.

“The main goal is to help human creativity,” said Oh, who is looking for ways to create more intuitive interactions between human and AI. People want to blame artificial intelligence. But the end product is our responsibility.”

This story was first published in the Recode newsletter. Register here So you don’t miss the next!

Leave a Comment